TOUCAN: Synthesizing 1.5M Tool-Agentic Data from Real-World MCP Environments
Learn how TOUCAN creates high‑quality, multi‑tool, multi‑turn data and why it beats larger closed models on benchmark tasks.
TOUCAN delivers 1.5 million agentic trajectories synthesized from nearly 500 real‑world Model Context Protocols, filling a glaring gap in open‑source tool‑agent training data. Existing collections lack the diversity and realism needed for multi‑tool, multi‑turn interactions, limiting progress in the field. By grounding generation in authentic MCP environments, TOUCAN produces tasks that mirror real tool execution. The scale and fidelity of the data set a new baseline for permissively licensed resources.
The pipeline begins by prompting five distinct language models to generate a wide range of tool‑use queries across the MCPs. Model‑based quality filters prune low‑quality prompts before three teacher models, each paired with two agentic frameworks, expand the queries into full trajectories. Each trajectory records the sequence of tool calls, inputs, and outputs, preserving the causal chain of execution. Rule‑based checks and additional model‑based validation enforce consistency, correct tool usage, and realistic state transitions.
Three extensions further diversify the dataset: they inject variations in tool parameters, introduce stochastic environmental changes, and simulate multi‑turn conversational contexts. These augmentations increase the difficulty of the tasks and expose agents to richer interaction patterns. The design deliberately balances breadth, covering many tools and protocols, with depth, ensuring each trajectory reflects a plausible real‑world scenario. The result is a heterogeneous corpus that challenges agents beyond single‑tool, single‑turn settings.
Models fine‑tuned on TOUCAN surpass larger closed‑source competitors on the BFCL V3 benchmark and achieve a new Pareto optimum on the MCP‑Universe benchmark. The abstract reports consistent gains across metrics, indicating that the synthetic trajectories translate into measurable performance improvements. The dataset’s scale and quality enable smaller open‑source models to close the gap with proprietary systems, demonstrating the practical impact of high‑fidelity tool‑agent data.
TakeawayOpen‑source agents can match or exceed closed models when trained on large, realistic tool‑use trajectories like those in TOUCAN.